Material-Aware Grasp Success - Dataset

A material-conditioned grasp-success benchmark 701,320 grasps × 19 materials = 13,325,080 hold/slip labels over 7,806 ShapeNetSem objects. Every grasp is simulated in MuJoCo (Warp backend) under all 19 materials, with geometry and pose held fixed. The material is therefore the sole causal driver of the outcome: a do(material) intervention. The primary evaluation slice is the decisive subset (~38.4% of grasps), those whose outcome flips across materials. Code and full method: https://github.com/collab-uniba/material-aware-grasping Files File Contents labels.npy Hold/slip labels, shape (701320, 19), int8 ∈ {0, 1}. Row i is grasp i; column m is material m, ordered as in materials.json. grasps.parquet 701,320 rows. Grasp keys (mesh_hash, scale, grasp_index) and contact-geometry features: V, logV, category, antipodal_min, antipodal_left, antipodal_right, grasp_width_mesh, width_slack, normal_disagreement, moment_arm. grasp_pose.parquet 701,320 rows. The 6-DoF grasp pose: position, approach and closing axes, 6-D rotation, quaternion, gripper width. Derived from ACRONYM. materials.json The 19 material names with their raw physics table (density, friction, restitution, hardness) and a z-scored version. physics_table.json Simulation physics parameters and metadata. folds.csv The frozen 10-fold assignment (mesh_hash to fold), shared by every model. SHA256SUMS.txt Checksums for all data files. All rows are aligned: row i of labels.npy, grasps.parquet and grasp_pose.parquet refer to the same grasp. The join key is (mesh_hash, scale, grasp_index). Decisive subset A grasp is decisive when its outcome flips across materials, i.e. labels[i].min() == 0 and labels[i].max() == 1. It holds under some materials and slips under others. This accounts for ~38.4% of grasps and is where the material question is empirically testable. The headline metric is AUROC on this subset. Materials (19) Metal, Ceramic, Glass, Granite, Stone, Concrete, Brick, Wood, Paper, Plastic, Wallpaper, Cardboard, Fur, Leather, Rubber, Foliage, Carpet, Fabric, Food. Point clouds and meshes (not included) The object point clouds and the ShapeNetSem meshes are not redistributed here, per the ShapeNet Terms of Use. Anyone holding ShapeNetSem can regenerate the point clouds from the mesh_hash and scale keys using the code repository. The labels, features, poses and splits provided here are sufficient to train and evaluate the tabular models and to reproduce the fold protocol. License and attribution Labels, features, materials and splits (our simulation outputs) are released under CC BY-NC 4.0. The grasp poses are derived from ACRONYM (Eppner et al., NVIDIA), also CC BY-NC 4.0, which requires attribution and permits non-commercial use only. The meshes are ShapeNetSem, subject to the ShapeNet Terms of Use, and are not included. The package as a whole is CC BY-NC 4.0. Citation Please cite this dataset together with: Eppner, C., Mousavian, A., Fox, D. ACRONYM: A Large-Scale Grasp Dataset Based on Simulation. ICRA 2021. Chang, A. X. et al. ShapeNet: An Information-Rich 3D Model Repository. arXiv:1512.03012, 2015.